Implement StridedArrayView

[bergbauer]

Needed for #15971

[bangerth]

Suggested change

	template <typename ElementType, int stride = 1>
	template <typename ElementType, std::size_t stride = 1>

[bangerth]

Can you move the definition out of line of the class declaration?

Also, please use the initializer list feature:

Suggested change

	StridedArrayView(value_type *starting_element, const std::size_t n_elements)
	{
	this->starting_element = starting_element;
	this->n_elements = n_elements;
	}
	StridedArrayView(value_type *starting_element, const std::size_t n_elements)
	: starting_element (starting_element)
	, n_elements (n_elements)
	{}

[bangerth]

Move the definition out of line. Same for the functions below.

[bergbauer]

@bangerth Updated!

[kronbichler]

I think this separation makes sense. I will not approve because I was involved in discussions. One thing I will mention is that we discussed that it is good to provide the pointer iterators begin and end only for the actual ArrayView class, because for a strided class one would need non-standard iterators, which is probably beyond the scope for a simple class like ArrayView.

[peterrum]

I am not sure if it is a good idea to make StridedArrayView a base class of ArrayView. StridedArrayView looks like a variant of a view; there are other views like in the case of slices of python.

I know why you want to do this. You want to simply extract the strides in FPE.

Personally, I would prefer to have StridedArrayView and ArrayView and accept both classes in FPE and decide internally via const expressions what kind of strides to use. This is just a bit more code in FPE.

[bergbauer]

Personally, I would prefer to have StridedArrayView and ArrayView and accept both classes in FPE and decide internally via const expressions what kind of strides to use. This is just a bit more code in FPE.

This would of course also be possible but I think StridedArrayView is a generalization where ArrayView is the special case with stride = 1. The question is if StridedArrayView can be useful in other parts of the library. If yes, I would go with the in this PR proposed solution, if no I can also implement a special solution for FPE. What do others think?

[peterrum]

StridedArrayView is a generalization where ArrayView is the special case with stride = 1. The question is if StridedArrayView can be useful in other parts of the library.

This is one point of view. The other one is a filtered vector, where all entries are "enabled".

[bergbauer]

Can we come to a decision here? I would like to move forward with #15971

[kronbichler]

The question is if StridedArrayView can be useful in other parts of the library

I do not see an immediate use, but one could imagine several potential uses in some parts (e.g. for tensor-product evaluation). But I think someone else should decide.

[kronbichler]

It seems there are no statements in support. Thus I suggest to solve the problem within FEPointEvaluation for now as suggested by @peterrum. We can close this for now but come back in case there is support for this PR. Does this make sense @bergbauer?

[bergbauer]

What about adding it as a parallel class (and not as base class of ArrayView, see last commit)?

Base class makes more sense for me because of the code duplication with this solution but this would also work.

[bangerth]

I don't have a particularly strong opinion about this patch, but would like to point out that we document ArrayView as comparable to the C++20 class std::span: https://en.cppreference.com/w/cpp/container/span . std::span explicitly mentions that it represents a contiguous range.

[kronbichler]

This looks ok for the present application.

[kronbichler]

/rebuild

Changes have been applied.

[peterrum]

Let's merge so that @bergbauer can proceed with #15971.